1. No category

Work-in-Progress
CHI 2015, Crossings, Seoul, Korea
CogniMouse: On Detecting Users’
Task Completion Difficulty through
Computer Mouse Interaction
Marios Belk
Eleni Christodoulou
Abstract
Dept. of Computer Science
CITARD Services Ltd.
University of Cyprus
CY-2064 Nicosia, Cyprus
CY-1678 Nicosia, Cyprus
cseleni@citard-serv.com
This paper presents a method for analyzing users’
computer mouse interaction data with the aim to
implicitly identify task completion difficulty while
interacting with a system. Computer mouse motion
streams and users’ skin conductance signals, acquired
via an in-house developed computer mouse, and users’
feedback were investigated as reactions to task
difficulty raising events. A classification algorithm was
developed, producing real-time user models of user
hesitation states. Preliminary results of a study in
progress with seven older adults at work (age 56+)
provide initial indications about links between mouse
triggering states of user hesitation and task completion
difficulty.
belk@cs.ucy.ac.cy
George Samaras
David Portugal
Dept. of Computer Science
CITARD Services Ltd.
University of Cyprus
CY-2064 Nicosia, Cyprus
CY-1678 Nicosia, Cyprus
davidsp@citard-serv.com
cssamara@cs.ucy.ac.cy
Permission to make digital or hard copies of part or all of this work for
personal or classroom use is granted without fee provided that copies are
not made or distributed for profit or commercial advantage and that
copies bear this notice and the full citation on the first page. Copyrights
for third-party components of this work must be honored. For all other
uses, contact the Owner/Author.
Copyright is held by the owner/author(s).
CHI'15 Extended Abstracts, Apr 18-23, 2015, Seoul, Republic of Korea
ACM 978-1-4503-3146-3/15/04.
http://dx.doi.org/10.1145/2702613.2732881
Author Keywords
Computer Mouse; Sensors; Older Adults; User Study.
ACM Classification Keywords
H.5.2. Information interfaces and presentation (e.g.,
HCI): User Interfaces.
Introduction
With the advent of highly dynamic and fast emerging
technologies and software, users are required to adapt
1019
Work-in-Progress
CHI 2015, Crossings, Seoul, Korea
their mental models, skills and habitual ways of
working. This requirement is further intensified, when
combined with eventual age-related cognitive
degradations of older adults [1], causing task
completion difficulty and eventually a negative user
experience.
Assisting older adults while interacting with systems is
of critical importance in today’s information society.
Researchers and practitioners alike have shown an
increased interest lately in understanding behavior
patterns and possible difficulties of older adults
imposed by current visual and interaction designs of
interactive systems [2, 3]. A number of research works
exist that proposed various intelligent user interfaces
and systems for supporting older adults at work and
motivating them to stay for longer active and
productive in computerized working environments [4,
5]. An important challenge of such intelligent and
assistive interactive systems is how to implicitly identify
that users experience difficulties with a given task [6],
and accordingly assist the users, e.g. by providing
personalized and contextualized support [7].
In this context, this paper presents an implicit user
data collection method for identifying users’ task
completion difficulty by leveraging computer mouse
motion streams and a skin conductance sensor that is
embedded in an in-house developed computer mouse.
This work is part of CogniWin project, which aims to
provide an innovative personalized system, motivating
older adults to stay for longer active, and improving
their productivity in the workplace. The paper is
structured as follows: next we present related works,
and subsequently we describe the developed computer
mouse, named CogniMouse. Afterwards, we present the
methodology and preliminary results of a user study
that evaluated the accuracy of the computer mouse’s
task difficulty identification method. We conclude with
practical implications and future prospects of this work.
Related Work
The literature proposes several systems that leverage
users’ computer mouse interaction data in order to
implicitly infer important information about the users’
behavior patterns. Principally, the raw mouse data
being extracted and processed includes time-stamped
and chronologically ordered sequences of a stream of
data including the x-y cursor position of the mouse on
the screen, mouse hover, mouse scrolling and mouse
click events (left and right clicks) [8-11].
Research works have proposed various approaches for
distinguishing users’ behavior patterns, such as user
hesitation, reading by tracing text, clicking and scrolling
[8, 9, 10]. Reeder and Maxion [8] proposed an
automated method for identifying user hesitation with
the aim to detect instances of user difficulty while
interacting with a system. Based on a hesitation
detector algorithm that takes as input a time-stamped,
chronologically ordered data stream of mouse and
keyboard events, user hesitation was defined as
anomalously long pauses between events in a given
data stream. Arroyo et al. [9] proposed a Web logging
system that tracks mouse movements in Websites
illustrating mouse trajectories of users’ interactions that
indicate users’ reading behavior and hesitation in the
menu area. A preliminary study of Mueller and Lockerd
[10] revealed that users often move the mouse cursor
to an empty white space of the Webpage in case they
are hesitating in order to avoid accidental clicks on
hyperlinks.
1020
Work-in-Progress
CHI 2015, Crossings, Seoul, Korea
Unlike the aforementioned works, we propose an
innovative instrumented computer mouse for advanced
user interface sensing; as well as a new technique for
detecting user hesitation with sensor fusion, which is
not solely based on input pauses, but instead on realtime computer mouse data analysis, which benefits
from detecting when the user is actually touching the
device. Moreover, we validate preliminarily our system
by conducting a user study, which indicates potential
links between user hesitation state and task completion
difficulty.
programming language, which has been chosen for
several reasons: i) it provides a good tradeoff between
efficiency, security and robustness; and ii) it provides a
number of libraries for communication with hardware
components, e.g. HID support.
User Hesitation Classification
Delay is a common consequence of users having
difficulty with a particular task [8]. Accordingly, the
developed method detects the level of difficulty in
completion of tasks by the user, through the detection
of hesitating behavior.
The CogniMouse Architecture
Figure 1. The CogniMouse
architecture.
To date, the main focus has been placed in the
development of a classifier of user’s hesitation state
according to the GSR sensor input, and mouse motion
streams. In this context, a parser module receives and
processes the raw data coming from the mouse
sensors, which is then analyzed by the classifier
algorithm.
We have implemented a Bayesian Classifier to detect a
general measurement of user’s hesitation. Different
classification techniques could be employed [12],
however the classification algorithm highly depends on
the properties of the specific problem to tackle, and the
application of classification methods in human state
prediction from multiple sources of data has different
number of features and challenges. Assuming the
availability of user historical data and common interval
of parameters which are deemed as usual for each
individual, human expertise judgment beforehand can
be leveraged to model relevant data, with an important
impact on the inference process. This is the case of
Bayesian inference, which makes sequential use of the
Bayes’ formula, when more data becomes available, to
calculate a posterior distribution. This provides means
to make inference about an environment of interest
described by a state given an observation, whose
relationship is encoded by a joint probability
distribution.
Following the philosophy described above, CogniMouse
is currently being developed using C# object-oriented
Specifically, our model represents the probability or the
level of certainty that the user is having difficulty in
CogniMouse is a plug-and play Human Interface Device
(HID) that connects to any host computer via a
Universal Serial Bus (USB) as seen in Figure 1. Having
driver support on all major operating systems, a
custom HID protocol has been developed, which allows
sending 64 bytes of different sensor data per packet.
On the mouse buttons area, a transparent Galvanic
Skin Response (GSR) sensor has been embedded to
react to changes on the user’s skin response. From the
outside, the current prototype design makes use of a
commercial Microsoft Comfort Mouse 4500 (cf. Figure
2), which causes no barrier to the users’ acceptance.
1021
Work-in-Progress
CHI 2015, Crossings, Seoul, Korea
completing the task, given by P(Hes|ZC,V,GSR). To
that end, we used three different inputs in our Bayesian
Classifier: i) the number of zero-crossings of the
motion vector over the last 4 seconds of mouse usage
(ZC); ii) the average intensity of the velocity vector
over the last 4 seconds of mouse usage (V); and iii) the
galvanic skin response value (GSR). The zero-crossings’
input represents the number of times that the user
changed motion direction when using the mouse.
Since we are developing a general model and no
personalized data of the user is available at this stage,
we assume that there is no prior information about the
user. Thus the prior distribution, P(Hes), is defined as
uniform, where all decisions are equiprobable.
Additionally, we need to define the likelihood functions,
P(input|Hes), so as to model each input according to
the expected hesitation.
The hesitation state of the user is obtained by fusing all
the inputs by applying Bayes Formula in each iteration
of the following algorithm:
Figure 2. The current prototype
design of the computer mouse
makes use of a commercial
Microsoft Comfort Mouse 4500.
|
, ,
∙
|
|
∙
∙
∙
|
∙
Method of Study
A user study was conducted with the aim to initially
evaluate the accuracy of the user hesitation
classification algorithm.
Procedure
Controlled laboratory sessions were conducted in which
participants were required to perform a series of tasks
(i.e., change settings of a Web browser) on a standard
desktop computer (IBM Thinkcenter M73, 21’’ monitor),
using a standard keyboard and the developed computer
mouse which recorded the mouse motion stream data
and users’ skin conductance signals. Screen capturing
software and audio data based on the think-aloud
protocol were recorded which were later analyzed by a
user experience researcher aiming to identify users’
task completion difficulty. This information was critical
since it serves as an evaluation basis for the user
hesitation triggering events of CogniMouse. A pre-study
questionnaire was also provided to the participants to
rate their experience with computers, the frequency of
using the Web browser in their daily work, and the
frequency they use Web browser tools, i.e. General
Settings, Privacy, Security and Advanced Settings.
Users’ Task Completion Difficulty Identification
Based on the screen capturing software and the thinkaloud audio data, users’ task completion difficulty while
performing a particular task was empirically detected
and evaluated based on the following criteria: i) user
statements (e.g. “I don’t know …”); ii) user silence and
inactivity of computer mouse and keyboard for a
considerable period of time; and iii) user asks for help
from the user experience researcher. Periods of task
completion difficulty are noted as soon as one of the
above criteria are met until the user clearly performs
an action (e.g. mouse click) indicating that the user is
not in a difficulty state anymore, or when the user
states that he/she is going to perform an action or that
the difficulty has been overcome (e.g. “OK, I will …”).
Tasks
All participants were required to complete three tasks
within a Web browser (Google Chrome v.39). The
following tasks were provided in a random sequence to
all participants, which primarily require changing the
1022
Work-in-Progress
CHI 2015, Crossings, Seoul, Korea
Web browser’s settings: i) “Set your preferred Website
as your Web browser’s start page”; ii) “Set the search
engine to be used when searching from the Web
browser’s search box (omnibox)”; iii) “Set your Web
browser’s cookie policy to block Websites from writing
any data on your computer”. These tasks were chosen
since they are performed by the users using primarily
the computer mouse, and require limited keyboard
typing input.
Figure 3. Means of hesitation values
per user and task.
Figure 4. Means of hesitation
values per task and user group.
Data Analysis
Two types of data were analyzed: i) user hesitation
output, triggered by the CogniMouse user hesitation
classification algorithm; and ii) task completion
difficulty as noted by the user experience researcher
based on the screen capturing software and the thinkaloud audio data analysis.
Participants
A total of seven individuals (4 male and 3 female)
participated in the study, and their age varied between
56 and 67 (mean age 62). All participants are using
computers and accessing Websites on a daily basis.
Based on the pre-study questionnaire, four participants
consider themselves as average to experienced, while
three consider themselves as novice to average in
regards with knowledge and experience with
computers. All participants use the Web browser for the
daily work activities. In regards with frequency of
handling the settings of the Web browser, three
participants rated themselves as experienced, whereas
the remaining four as novice.
Analysis of Results
For our analysis we have separated users into two
groups: the High-difficulty Group that includes four
participants that had significant task completion
difficulties while performing all the three tasks, and the
Low-difficulty Group that includes three participants
that did not have any major task completion difficulties
while performing the tasks. The categorization into
these groups was based solely on the empirical analysis
made based on the screen capture data and the thinkaloud audio analysis. Figure 3 illustrates the means of
user hesitation values per user and task; High-difficulty
Group includes User #1-4, and the Low-difficulty Group
includes User #5-7. Figure 4 illustrates the means of
user hesitation values per task and user group.
An independent-samples t-test was run to determine if
there were differences in user hesitation values
between the High-difficulty and Low-difficulty user
groups. There was homogeneity of variances, as
assessed by Levene's test for equality of variances
(p=0.475). Results revealed that users of the Highdifficulty Group triggered higher user hesitation scores
(0.65±0.045) than those of the Low-difficulty Group
(0.57±0.026). These were statistically significant
different at 0.082 (t(5)=2.757, p=0.04, d=0.21).
Furthermore, descriptive statistics reveal that users of
the Low-difficulty Group scored lower user hesitation
values in all three tasks (Figure 4).
Conclusions and Future Work
The purpose of this paper is to present results of a
work in progress that aims to implicitly identify users’
task completion difficulty by leveraging computer
mouse motion data and users’ skin conductance
signals. A user study was conducted with seven older
adults to investigate the accuracy of a user hesitation
classification algorithm. Empirical data, based on
screen capture information and think-aloud audio data,
1023
Work-in-Progress
CHI 2015, Crossings, Seoul, Korea
revealed that users having difficulty with particular
tasks scored higher user hesitation values compared to
users that did not face any major task completion
difficulties. Although the analysis yielded statistical
significant results, these results shall be confirmed in
further studies with a larger sample and users with
varying profiles and ages. Finally, we believe that even
though this study is focused on older adults, in general,
identifying task completion difficulty through the
computer mouse could be suitable for other age
groups.
In the future, we intend to contextualize the mouse
data so that the system provides adaptive support
taking the user’s task into account, when difficulty is
identified. Moreover, further sensors shall be embedded
in CogniMouse to detect additional user behaviors, as
well as information from other devices, such as an eye
tracker, which will be integrated in order to trigger
personalized assistance.
Acknowledgements
This work was partially carried out in the frame of the
CogniWin project, funded by the EU Ambient Assisted
Living Joint Program (AAL 2013-6-114) as well as the
EU FP7 project Miraculous-Life (#611421) and the
Cyprus Research Promotion Foundation project
PersonaWeb (ΤΠΕ/ΠΛΗΡΟ/0311(ΒΙΕ)/10).
References
[1] Kausler, D.H., Salthouse, T.A., Saults, J.S.
Temporal memory over the adult lifespan. The
American Journal of Psychology 101, 2, 207-215.
[2] Findlater, L., Froehlich, J., Fattal, K., Wobbrock, J.,
Dastyar, T. Age-related differences in performance with
touchscreens compared to traditional mouse input. In
Proc. CHI 2013, ACM Press (2013), 343-346.
[3] Waycott, J., Vetere, F., Pedell, S. Designing
communication technologies for socially isolated older
people: Fieldwork reflections. In Proc. CHI 2013,
Workshop on Designing for -and with- Vulnerable
People, ACM Press.
[4] Leung, R., Findlater, L., McGrenere, J., Graf, P.,
Yang, J. Multi-layered interfaces to improve older
adults’ initial learnability of mobile applications. ACM
Trans. Access. Comput. 3, 1, Article 1 (2010), 30 p.
[5] Lindsay, S., Jackson, D., Schofield, G., Olivier, P.
Engaging older people using participatory design. In
Proc. CHI 2012, ACM Press (2012), 1199-1208.
[6] Atterer, R., Wnuk, M., Schmidt, A. Knowing the
user’s every move: user activity tracking for website
usability evaluation and implicit interaction. In Proc.
WWW 2006, ACM Press (2006), 203-212.
[7] Brusilovsky, P., Kobsa, A., Nejdl, W. The adaptive
web: methods and strategies of web personalization.
Springer Verlag, Heidelberg, 2007.
[8] Reeder, R. Maxion, R. User interface defect
detection by hesitation analysis. In Proc. DSN 2006,
IEEE Computer Society (2006), 61-72.
[9] Arroyo, E., Selker, T., Wei, W. Usability tool for
analysis of web designs using mouse tracks. Ext.
Abstracts CHI 2006, ACM Press (2006), 484–489.
[10] Mueller, F., Lockerd, A. Cheese: Tracking mouse
movement activity on websites, a tool for user
modeling. Ext. Abstracts CHI 01, ACM (2001), 279-280.
[11] Huang, J., White, R., Dumais, S. No clicks, no
problem: using cursor movements to understand and
improve search. In Proc. CHI 2011, ACM Press (2011),
1225-1234.
[12] Luz, J.M., Couceiro, M., Portugal, D., Rocha, R.P.,
Araújo, H., Dias, G. Comparison of classification
methods for golf putting performance analysis. Comp.
Intel. and Dec. Mak. 61, 35-45.
1024